The number of these things is the numerator of a ratio that is empirically optimal at about 20, the Chinchilla point. Computations involving the Q, K, and V vectors generated from these things are stored in a KV cache, (-5[3])which improves a metric (-5[1])called the “time to the first” one of these things. (-5[3])The process of converting an input into these (10[1])things is commonly done (-5[1])using byte-pair encoding. (10[1]-5[1])During (10[1]-5[1])the prefill stage, (10[2])IDs corresponding to these things (-5[1])are embedded into a vector that is then passed through multiple transformer layers. (-5[2])The number of these (10[1])things (10[1])that can be processed at once is the context (10[1])window. (10[3]-5[2])The (10[3])cost (-5[1])of (10[3])AI (10[2])compute (10[3])is often (-5[1])reported (10[1]-5[1])as (-5[1])the price (10[2])per (10[1])one million of these (10[1])things. (-5[5])For (10[1])10 points, (10[1])name these units of text (10[2])that (10[1])large language models (10[1])split queries into and try (10[1])to predict the next one of. (0[1])■END■ (10[13]0[13])